Improving Credit Decisioning Models

Credit decisioning is the process of making a decision to extend credit facility to a counterpart. It is a core business for banks and the profitability depends on the accuracy of decisioning. It is a complex process which takes into consideration multiple interdependent elements.

Gini Coefficient of a model describes the effectiveness of a model in differentiating between the good borrowers and bad borrowers. It is also used to compare models on their predictive capability. Objective is to improve the Gini score of the model by improving the predictability. This is done by reducing the number of false positives and false negatives. False positives are those cases where model predicted a default and borrower didn't default. False negatives are those cases where model predicted no default and borrower defaulted.

3 Benefits which banks can reap from using high-performance credit decisioning models

1. Increase in revenue - Better distinguishing between credit worthy and non credit worthy customers, helps banks improve acceptance rates and pricing.
2. Reduction in credit-loss rate - Decrease of 20 to 40 percent in credit loss by using better models for predicting customer default. This affects provisions which banks must hold
3. Efficiency gains - Automated data extraction, case prioritisation (for example straight through processing for low risk cases while analysing high risk cases more thoroughly) and model development.

As customer information becomes more democratised via Open Banking and regulations such as PSD2 and as fintech companies and attacker banks proliferate and focus on increasingly digitally savvy customer base, it becomes increasingly difficult for incumbent banks to preserve market share and profitability.

Banks can improve decisioning by having Credit decision models that can include new data points, model accurate customer behaviours, identify new segments and react faster to business environment changes.

Ways of improving credit decision models

• Modular Architecture - By using a modular structure, new data sources and new information can be added and removed as modules making the structure robust and flexible. The submodes scores can be aggregated using different weight sensitivities to form one final score.

For Example a pandemic submodule can be created which can take into consideration the change in cash and net income to signal financial distress in the model. Similarly how the management is treating a business is also a good indicator of future solvency of the company specially in the SME space.

Following a customer-centric approach vs a product-specific approach leads to a higher performance model as customer centric approach takes into consideration data signals related to all product areas with which a customer interacts. Model development teams must validate assumptions with business. The data being used should be distinct as data overlap can skew the results. Financial data sources from two sources can result in double counting of financial factor thus over emphasising the impact.

• Expanding data sources - Traditionally banks use only internal data sources for credit modelling. But supplementing the internal data source with non traditional and external traditional and non traditional data sources improves the model. For example using customer behaviour data from access devices such as mobile phones, tablets or social media etc can bring additional data points for consideration.

Open banking is being leveraged for adding transactional data. By accessing data across banks Open banking makes it possible to build a complete picture of income and expenses based on the transactions data. Telecom data of bill payment history, Social media data on professional and personal travel, connections, jobs etc are also example of non traditional data which can improve the model.

• Data mining on existing data - Banks have huge amounts of data which is often not used optimally to generate credit signals. Some banks use open banking data to analyse income and spend patterns and design synthetic financial and cash flow statements for their customers to identify credit signals and segment customers. Some banks use text-mining and natural language processing to tag transactions more accurately and use this enhanced personal and business transaction data for credit modeling.

Local factors are more important for SME businesses, whereas risk factors derived from company reports are more relevant for large businesses. ML techniques can be used to identify and define new customer segments. ML can be used to develop challenger models to identify other signals which can improve the incumbent model score.

• Leveraging business expertise - Banks must use internal expertise to augment the credit models, RMs often have a final say in the decisioning but their inputs are also very valuable in developing the models. Underwrites and RMs can point to specific areas which can be used to define credit signals. They can also leverage their knowledge of business, bank policies and compliance procedures to emphasise the importance of certain inputs at specific time periods.

Ref -

1. using-the-gini-coefficient-to-evaluate-the-performance-of-credit-score-models-59fe13ef420
2. https://www.fico.com/blogs/breaking-down-origination-decision-model
3. https://www.mckinsey.com/capabilities/risk-and-resilience/our-insights/designing-next-generation-credit-decisioning-models